2026-07-24 · Notes

Harvey's ROI Math Is Right. Its Finnish Context Is Missing.

Harvey just published its Legal AI ROI Guide. The numbers are bold: typical users save 15–25 hours a month, power users 30+, and the sample calculation shows a 20x return on the license. The core thesis is right, though. Value comes from giving the model full context on a lawyer’s work.

Which is exactly why the pitch falls short in Finland. Full context here means Finlex: current statutes, Supreme Court and Supreme Administrative Court precedents, and government bills. Neither Harvey nor Legora currently offers an open MCP connection to the Finnish legal sources discussed here. They build a closed harness between the model and the user, and price the harness as if it were the expertise itself.

Laki.ai took the opposite route and shipped an open MCP server. Any frontier model can now query the interlinked Finlex corpus directly: open a section of law and the related bills and case law are right there.

On the latest Neuvottelija episode, Aleksi Paavola, who built the Laki.ai MCP, compares Harvey and Legora to Cursor before the model leap. The labs keep stripping away the harness, and Claude connected to Laki.ai plus a Finnish legal skills library can already provide stronger results on many Finnish legal tasks than closed wrappers with limited knowledge of Finnish legal sources.

So the ROI logic cuts both ways: before paying for legal AI, ask which legal sources it can actually reach in your jurisdiction.

The longer version

That summary compresses a seventy-five-minute conversation into one sentence, and the argument deserves more room — including the parts that cut against my own conclusion. What follows is the detailed reading of my conversation with Paavola (full episode).

The Bitter Lesson is the actual argument

The Cursor comparison is not an analogy about user interfaces. It rests on Rich Sutton’s essay The Bitter Lesson, which Paavola brings in explicitly (18:44): the value a human adds to an AI system tends to shrink over time, and the systems that win are the ones that let the model do more rather than less.

His two examples are the ones that make it concrete. In chess, people assumed a human could improve an engine by adding structure around it. The historical answer is unambiguous — the engine alone is stronger than engine plus human. In self-driving, the early approach was piecewise sub-optimisation, building small hand-specified pieces. Paavola points to Andrej Karpathy’s account of Tesla’s strategy — training the network to drive end to end — turning out better than steering the model with human intelligence at particular points (19:46).

Then software development, which is where the pattern becomes recent enough to feel. Cursor took a traditional IDE and integrated AI into it, and for a long while tab completion looked like the shape of the future: you write the start of a function, the model predicts the rest. Paavola admits he was sceptical that Claude Code’s plain text interface could work at all — and then notes he cannot remember the last time he used tab completion (20:51).

The load-bearing observation is what the labs themselves say:

If we look at what OpenAI’s or Anthropic’s employees say, what they talk about is exactly that they’re constantly trying to strip that harness down — letting the model take on a bigger and bigger share of it. (21:55)

Which places Legora and Harvey structurally, not rhetorically. They have built software around the legal task so as to add value at specific points. That is the pre-leap Cursor position, and the direction of travel is against it.

The Finnish gap is specific, not general

This is the part that turns a thesis about industry direction into a purchasing decision, and it is sharper than my original note conveyed.

They have the capability — in Sweden. Legora acquired a Swedish startup that had done for Swedish law roughly what Laki.ai has done for Finnish law. So the capability exists inside the product; it simply does not extend to Finland (26:03). This is not a technology gap. It is a market-size decision, and Finland lost it.

So what does Finnish legal research inside those products actually run on? A web browser. And here Paavola lands the observation that should give any Finnish buyer pause (27:06):

A good example is Harvey’s own case, where by default they’ve actually banned the model from doing this web search — and there’s a very good reason for that. And now you’ve got this appallingly expensive product whose legal research relies on that web search.

Read that twice. The vendor’s own default setting is an admission about the reliability of the channel, and in Finland that channel is the primary one.

And you cannot patch it yourself. As far as Paavola knows, you cannot connect Laki.ai into Legora or Harvey. The harness is closed in both directions: it does not reach Finnish sources, and it does not let you reach them either.

What is actually in the stack that works

Worth being concrete about what replaces it, because “just use Claude” is not the recommendation.

MCP is Anthropic’s open protocol, and Paavola’s analogy for it is USB-C: a standard port so an agent can plug into an application without either side building something bespoke (09:21). Underneath it there are still ordinary APIs; MCP is the layer that makes them usable by an agent rather than by a developer.

What Laki.ai adds beyond Finlex itself is the part I underplayed. Finlex has the statutes, the case law and the government bills. Laki.ai’s contribution is that the documents are linked to one another (06:14): open a given section of law and you immediately see which bills it traces to and which court cases attach to it. For a research agent that is the difference between a library and a citation graph.

The skills layer is the second half. Aku Nikkola of Legit published a Finnish adaptation of Claude’s legal skills — Claude’s own pack encodes Anglo-Saxon practice, and his team built one for the particulars of Finnish law, starting from the language and the structure of a statutory section (07:17, 31:16). Combined with the Laki.ai MCP, that is the working Finnish stack today. It is two open components, not one product.

A practical finding on building skills, which generalises well beyond law: I had been building them alongside the work — watch me do this task, note what comes up. That is backwards. Do the whole task through to the finish line first, then turn the finished path into a skill (31:16). Paavola agreed, and both of us noted there is no Stack Overflow for this any more; everyone is working it out at random.

And a workflow that beat either component alone. Running a draft through a local open-weights model produced nothing usable; a stronger local model was better but still thin. What worked was a rough base draft from the best available frontier models, then refinement through Laki.ai and the Finnish skills pack (32:19, 33:22). The reason two models help is that they have to be different models — the same model arguing with a copy of itself produces no real cross-examination.

Where the closed harness still wins

My original note did not carry this, and it should. There are three places where the argument I made does not reach.

Genuinely multi-dimensional work. I put this to Paavola directly (23:59): editing a straightforward legal document is one thing, but a multi-year tangle of shareholder agreements across five family offices’ portfolio companies is another. Vertical, task-by-task work is being chewed through continuously. Holistic work across many simultaneous dimensions is not the same problem, and there a thick harness still carries value — until enough of the verticals are solved that the system can pull them together itself.

Proprietary corpora that no open MCP can reach. Finland has weak courts and a strong Chamber of Commerce arbitration system (28:07). Arbitral awards bind the parties like a judgment, and they are not public. That body of knowledge sits in law firms’ own databases, because those firms were parties — and their partners may have been the arbitrators. A top Finnish firm’s internal archive of thousands of past matters is real, case-specific expertise, and open Finlex access does nothing to replicate it. The competitive question for such a firm is not Harvey versus Claude; it is whether it can get its own corpus into whichever model it uses.

The tooling got harder, not easier. Paavola’s own expectation a couple of years ago was that general-purpose AI would become simple enough for anyone. It went the other way (29:10). Law firms do not traditionally carry deep IT capability, let alone AI capability — so the bottleneck moves from the model to the integration, and a closed product that someone else operates is a legitimate answer to a staffing problem. That is worth paying something for. It is not worth paying for as though it were the expertise.

The one dimension they still own, and its expiry date

Interface. Harvey and Legora have UI components that a chat window does not, and that is a real advantage today. Paavola’s prediction is that it is temporary, because generative interfaces are arriving quickly: you describe the use case and the model builds the interface for it on the fly (25:02).

If that holds, the remaining room gets tight — and the commercial exposure is asymmetric. These products have been sold extremely well, at a steep price, on long contract terms (26:03). A firm that has built its whole process around one of them is committed for years. A smaller firm that adopts an open stack is not committed to anything.

The caveats I would attach before anyone acts on this

Anonymise before uploading. When you put confidential documents into any such system, pseudonymise the names — replace a client with “Person A1” and convert back after the run — particularly for genuinely sensitive matters such as inside information on listed companies (70:36). An anonymisation wrapper is the obvious missing open-source component here.

The target is that your licence becomes the limit. Paavola’s stated direction for Laki.ai is that the constraint should be your Claude or ChatGPT plan: on a Business or Enterprise tier you could put in whatever you would already put in email or OneDrive (72:42). That is a reasonable bar, and it still requires trusting the model vendor — a judgement each organisation has to make for itself.

EU-server requirements are a real exception. If your use case requires data residency in the EU, note that the model vendor’s cloud may not be. Read the fine print rather than assume (73:44). EU legislation is coming into Laki.ai, and the AI Act’s final implementation is close, with the Omnibus having pushed some of the worst of it further out.

The ROI number that closes the loop

Harvey’s guide computes a return against its own licence fee. The comparison a Finnish buyer should actually run is different.

The commercial equivalent of what Laki.ai provides — reliable, interlinked access to Finnish legal sources — costs a consumer several thousand euros a year from the incumbent publisher. Laki.ai is free (70:36). For an individual with a legal problem, several thousand euros is not a realistic price for the ability to read the law that governs them.

So the ROI question is not what multiple do I get on the licence. It is:

1. Which Finnish legal sources can this product actually reach, and through what channel?
2. What am I paying for the harness, as distinct from the model underneath it and the sources beside it?
3. What is my own corpus worth, and can I get it into whichever system I choose?
4. How long is the contract, given that the harness is the part shrinking fastest?

Harvey’s arithmetic is fine. It is the denominator that is jurisdiction-specific.

Watch the full episode on YouTube.

← All notes · Markdown mirror: index.md